Improved intelligent methods for power transformer fault diagnosis based on tree ensemble learning and multiple feature vector analysis

Abstract

Abstract This paper discusses the impact of the feature input vector on the performance of DGA-based intelligent power transformer fault diagnosis methods. For this purpose, 22 feature vectors from traditional diagnostic methods were used as feature input vectors for four tree-based ensemble algorithms, namely random forest (RF), tree ensemble (TE), gradient boosted tree (GBT), and extreme gradient tree (XGB). To build the proposed diagnostics models, 407 samples were used for training and testing. For validation and comparison with the existing methods of literature 89 samples were used. Based on the results obtained on the training and testing datasets, the best performance was achieved with feature vector 16, which consists of the gas ratios of Rogers’ four ratios method and the three ratios technique. The test accuracies based on these vectors are 98.37, 96.75, 95.93, and 97.56% for the RF, TE, GBT, and XGB algorithms, respectively. Furthermore, the performance of the methods based on best input feature were evaluated and compared with other methods of literature such as Duval Triangle, modified Rogers’ four ratios method, combined technique, three ratios technique, Gouda triangle, IEC 60599, NBR 7274, clustering, and key gases with gas ratio methods. On validating dataset, diagnostic accuracies of 92.13, 91.01, 89.89, and 91.01% were achieved by the RF, TE, GBT, and XGBoost models, respectively. These diagnostic accuracies are higher than 83.15 % of the clustering method and 82.02 % of combined technique which are the best existing methods. Even if the performance of DGA-based intelligent methods depends strongly on the shape of the feature vector used, this study provides scholars with a tool for choosing the feature vector to use when implementing these methods.